python-voiceio 0.9.8__tar.gz → 0.9.9__tar.gz

{python_voiceio-0.9.8/python_voiceio.egg-info → python_voiceio-0.9.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-voiceio
-Version: 0.9.8
+Version: 0.9.9
 Summary: Voice dictation for Linux. Speak → text, locally, instantly.
 Author: Hugo Montenegro
 License-Expression: MIT

{python_voiceio-0.9.8 → python_voiceio-0.9.9}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "python-voiceio"
-version = "0.9.8"
+version = "0.9.9"
 description = "Voice dictation for Linux. Speak → text, locally, instantly."
 readme = "README.md"
 license = "MIT"

{python_voiceio-0.9.8 → python_voiceio-0.9.9/python_voiceio.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-voiceio
-Version: 0.9.8
+Version: 0.9.9
 Summary: Voice dictation for Linux. Speak → text, locally, instantly.
 Author: Hugo Montenegro
 License-Expression: MIT

{python_voiceio-0.9.8 → python_voiceio-0.9.9}/tests/test_postcorrect.py

@@ -121,23 +121,15 @@ def test_disfluency_mode_rejects_rewording():
         assert pc.correct(original) == original
 
 
-def test_disfluency_mode_rejects_dropped_negation():
-    """Dropping 'not' is one deletion under every fraction cap but inverts
-    meaning — the negation guard must reject it."""
+def test_disfluency_mode_allows_trailing_shrug_with_negation():
+    """No mechanical negation guard — the LLM decides. Dropping a trailing
+    'I don't know' shrug (plus other filler) is accepted; the model, not a
+    regex, is trusted to avoid genuine polarity flips (see the prompt)."""
     pc = PostCorrector(_cfg(remove_disfluencies=True))
-    original = "I do not want the feature to ship today at all"
-    inverted = "I do want the feature to ship today at all"
-    with patch("voiceio.llm_api.chat", return_value=inverted):
-        assert pc.correct(original) == original
-
-
-def test_disfluency_mode_rejects_replaced_negation():
-    """Replacing a negation with something else also inverts meaning."""
-    pc = PostCorrector(_cfg(remove_disfluencies=True))
-    original = "we should never deploy on a friday afternoon"
-    changed = "we should always deploy on a friday afternoon"
-    with patch("voiceio.llm_api.chat", return_value=changed):
-        assert pc.correct(original) == original
+    original = "the build passed but I don't know we should still check it carefully today"
+    cleaned = "the build passed we should still check it carefully today"
+    with patch("voiceio.llm_api.chat", return_value=cleaned):
+        assert pc.correct(original) == cleaned
 
 
 def test_fix_mode_still_rejects_big_deletion():

python_voiceio-0.9.9/voiceio/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.9.9"

{python_voiceio-0.9.8 → python_voiceio-0.9.9}/voiceio/postcorrect.py

@@ -34,18 +34,17 @@ _MAX_INSERTED_WORDS = 0     # adding any word = altering meaning → reject
 _MAX_REPLACE_FRAC = 0.15    # ASR word-fixes only, never wholesale rewording
 _MAX_DELETE_FRAC = 0.4      # backstop against deleting real content
 
-# Words whose deletion/replacement flips meaning — the fraction caps can't catch
-# a single dropped "not". If the edit touches any of these on the original side,
-# reject outright. (Contractions ending in "n't" are handled separately.)
-_MEANING_CRITICAL = frozenset({
-    "not", "no", "never", "none", "nor", "neither", "without", "cannot",
-    "nothing", "nobody", "nowhere", "n't",
-})
 
+def _norm_words(text: str) -> list[str]:
+    """Lowercased, punctuation-stripped word list for the guard's diff, so a
+    sentence split or capitalization isn't misread as a content change."""
+    out = []
+    for w in text.split():
+        t = w.lower().strip(".,;:!?\"'()[]{}—–-")
+        if t:
+            out.append(t)
+    return out
 
-def _is_meaning_critical(word: str) -> bool:
-    w = word.lower().strip(".,;:!?\"'()")
-    return w in _MEANING_CRITICAL or w.endswith("n't")
 
 _SYSTEM_PROMPT = (
     "You fix automatic speech recognition errors in dictated text. "
@@ -74,10 +73,12 @@ _SYSTEM_PROMPT_CLEAN = (
     "   - stray word and phrase repetitions.\n"
     "Lean toward the briefer version — remove hesitation and padding freely.\n"
     "HARD LIMITS (these protect meaning): only DELETE and FIX — NEVER add "
-    "words, rephrase, reword, reorder, or summarize. NEVER drop or alter actual "
-    "content, a negation ('not', 'never', \"n't\"), or a hedge that changes "
-    "certainty ('I think', 'maybe', 'probably', 'might'). Keep the speaker's own "
-    "wording. Return only the cleaned text, nothing else."
+    "words, rephrase, reword, reorder, or summarize. NEVER change what a "
+    "statement asserts — above all never flip its polarity (do NOT turn 'I "
+    "don't think we should ship' into 'we should ship'), and keep hedges that "
+    "carry real uncertainty about a claim ('maybe', 'probably'). A trailing "
+    "verbal shrug like 'I don't know' with nothing after it is filler you may "
+    "drop. Keep the speaker's own wording. Return only the cleaned text."
 )
 
 _MAX_RECENT = 3
@@ -314,11 +315,13 @@ class PostCorrector:
     def _guard(self, text: str, corrected: str) -> tuple[bool, str]:
         """Decide whether the LLM's edit is within bounds.
 
-        Returns (accept, reject_reason). In disfluency mode the promise is
-        "never change meaning", enforced on the word-level diff: zero
-        insertions (nothing added/rephrased), few replacements (ASR fixes, not
-        rewording), bounded deletions (can't strip real content). Otherwise the
-        original conservative fix-only guards apply.
+        Returns (accept, reject_reason). In disfluency mode the LLM is trusted
+        to judge what's filler; the diff only backstops gross failure — nothing
+        ADDED (no rewrite/hallucination) and not more than a bounded fraction
+        deleted or word-swapped. Comparison is on normalized tokens (lowercased,
+        punctuation stripped) so cosmetic restructuring — a sentence split, a
+        capital letter, a comma — is never miscounted as rewording. Outside
+        disfluency mode the original conservative fix-only guards apply.
         """
         aw, bw = text.split(), corrected.split()
         if not self._remove_disfluencies:
@@ -329,18 +332,16 @@ class PostCorrector:
                 return False, "editratio"
             return True, ""
 
+        na, nb = _norm_words(text), _norm_words(corrected)
+        aw, bw = na, nb
         inserted = replaced = deleted = 0
         for tag, i1, i2, j1, j2 in difflib.SequenceMatcher(a=aw, b=bw).get_opcodes():
             if tag == "insert":
                 inserted += j2 - j1
             elif tag == "replace":
                 replaced += max(i2 - i1, j2 - j1)
-                if any(_is_meaning_critical(w) for w in aw[i1:i2]):
-                    return False, "negation"   # e.g. "not" → something else
             elif tag == "delete":
                 deleted += i2 - i1
-                if any(_is_meaning_critical(w) for w in aw[i1:i2]):
-                    return False, "negation"   # dropping "not" inverts meaning
         n = len(aw)
         if inserted > _MAX_INSERTED_WORDS:
             return False, "inserted"          # added content — meaning changed

python_voiceio-0.9.8/voiceio/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.9.8"